The purposes of this document

Halide already has an excellent tutorial that explains the syntax of the language and the basics of scheduling in detail. This document is an alternative to the official tutorial, that serves several complementary purposes:

• We provide a more lightweight introduction to Halide emphasizing what it can do and how to do it, with less focus on the scheduling language. We assume the autoscheduler can provide an efficient schedule.

• We show real Halide code examples and explain them, so that the readers can have a better understanding of Halide’s capabilities. We also briefly discuss strategies for manual scheduling.

• We show how to incorporate learning components in Halide programs with automatic differentiation, and how Halide interacts with deep learning frameworks such as PyTorch.